What the research actually says.

High-density EEG (256 channels) recorded 17 young adults as they drew pictures with a stylus, typed words on a keyboard, and described them aloud — comparing the brain activity each task produced.

Drawing produced theta- and alpha-band desynchronization across parietal and occipital regions — a pattern tied to memory encoding and learning. Typing produced no comparable engagement.

A 256-channel EEG study compared cursive handwriting with a digital pen, typing, and drawing in 12 university students and 12 seventh-graders, analysing the first seconds of each task.

Handwriting and drawing elicited theta-band synchronization in parietal and central regions — linked to memory and encoding — that typing did not. Children showed the same direction of effect as adults.

256-channel EEG measured brain connectivity in 36 university students as they handwrote words with a digital pen versus typed them with one finger.

Handwriting produced markedly more widespread theta/alpha coherence across parietal and central regions — the kind of connectivity associated with working memory and encoding — than typing.

Across three experiments, college students took notes by longhand or laptop during lectures, then were tested on factual recall and conceptual application.

Laptop note-takers transcribed more words verbatim and did worse on conceptual questions; longhand writers, forced to summarise, understood more. Even telling laptop users not to transcribe didn't help.

Preschool children were trained over three weeks to copy letters either by hand or on a keyboard, then tested on how well they recognised those letters.

Handwriting training led to better letter recognition than typing — the advantage clearest in the older preschoolers, who built richer motor-perceptual representations of each form.

Pre-literate children (~4–5 yrs) practised letters by free-form printing, typing, or tracing, then viewed letters during fMRI.

The brain's “reading circuit” — left fusiform gyrus, inferior frontal gyrus, posterior parietal cortex — was recruited during letter perception only after free-form printing, not after typing or tracing.

Adults learned novel letter-like symbols by writing them, typing them, or only viewing them; fMRI then measured the brain's response to those symbols.

Only after writing practice did activation to the new symbols come to resemble the response for real letters, notably in the left fusiform gyrus. Typing and viewing did not.

Students studied worked physics examples while thinking aloud; their explanations were analysed in detail and related to later problem-solving success.

Stronger learners spontaneously generated “self-explanations” that linked each step to underlying principles and monitored their own understanding; weaker learners mostly re-read the examples.

Children learning mathematical equivalence were crossed on prompted self-explanation (or not) and instruction type, then tested on retention and transfer to novel problems.

Prompted self-explanation improved transfer across conditions, and worked best combined with direct instruction.

A meta-analysis synthesising experimental studies of prompted self-explanation in mathematics across procedural knowledge, conceptual knowledge, and procedural transfer.

Prompted self-explanation gave small-to-moderate gains on all three immediate outcomes, strongest when explanations were scaffolded toward quality.

Seventh-graders learned to solve equations either by comparing two solution methods side-by-side or by studying the same methods one at a time.

Comparing methods produced greater procedural knowledge and flexibility than sequential study, with comparable gains in conceptual knowledge.

A follow-up varied the type of comparison: different solution methods, different problem types, or surface-equivalent problems.

Comparing different solution methods best supported conceptual knowledge and procedural flexibility; comparing surface-equivalent problems helped least.

In a real Harvard physics course, 316 students were randomized to a custom, pedagogically-constrained AI tutor or a high-quality in-class active-learning lesson on the same topics.

The AI-tutored group learned more than twice as much, in less time, and reported higher engagement and motivation (post-test medians 4.5 vs 3.5; p < 10⁻⁸). It was engineered with sequencing, scaffolding, and feedback — not raw ChatGPT.

The team built a dataset of 1,002 teacher-annotated math solutions, each marked at the first wrong step, and a tutoring pipeline that verifies the student's working before responding.

Grounding the tutor in a step-verifier produced more targeted feedback — more often correct, with fewer hallucinations than strong baselines — by locating the specific step where reasoning broke.

A meta-analysis of 53 studies (166 comparisons) testing whether having students attempt a hard problem before instruction beats the usual instruct-first order.

Problem-solving-before-instruction won overall (g = 0.36), and more strongly when it followed “productive failure” design principles.

A review synthesizing a century of evidence on two strategies: spacing study out over time, and retrieval practice — actively recalling material rather than re-reading it.

Both are among the most robust, broadly applicable ways to make learning durable — and both are badly underused, because students' own intuitions mislead them about what actually works.

Over three months, 126 seventh-graders did the same problems either blocked (one type at a time) or interleaved (types mixed, so they must choose which strategy applies), then took an unannounced test.

Interleaving won decisively, and the gap grew over time: 74% versus 42% correct on a test 30 days later (d = 0.79).

A randomized trial of a fully-online, small-group (two-to-one) math tutoring program for disadvantaged secondary students, run intensively over eight weeks.

It raised test scores by +0.26 SD and end-of-year grades by +0.49 SD, and cut the odds of repeating the year — at a fraction of the cost of one-to-one tutoring.

A landmark meta-analysis of 225 studies comparing active learning with traditional lecturing across undergraduate science, engineering, and mathematics.

Active learning raised exam scores by about 0.47 SD, and students in lecture sections were roughly 1.5× more likely to fail the course.